The Buyers' Strike: What is really driving the global bond sell-off

I have argued that the market had stopped taking AI capex on trust and started demanding a return on it. That argument was about equities. It has since migrated into the bond market, and it has arrived at an awkward moment. On 18 August the US 30-year Treasury yield touched 5.336%, its highest since 2007. French borrowing costs hit their highest since 2008, the 30-year Bund its highest since 2011, the 30-year gilt is flirting with 6% and the Japanese 40-year set a record 4.355% back in May.

This is not a country-specific fiscal accident. It is a synchronised repricing of long-dated debt, and the striking feature is that it is happening almost exclusively at the long end: the US 10-year, at around 4.74%, has moved far less. Jonas Goltermann of Capital Economics gave the consensus its cleanest formulation, saying the market is ‘demanding higher compensation for holding long-dated debt’. The question worth asking is compensation for what. Below are the five explanations currently doing the rounds, and the strongest argument against each.

The fiscal story

The obvious answer. US federal debt is closing in on $40trn. Moody's stripped the country of its last AAA in May 2025, projecting debt at 134% of GDP by 2035 against 98% in 2024, with deficits approaching 9% of GDP. France has been downgraded three times in twelve months, with Fitch putting its debt at 121% of GDP by 2027. The UK Debt Management Office is running a £252.1bn gilt remit for 2026/27. The bond vigilantes, we are told, are back.

However, deficits are a slow-moving variable and cannot on their own explain a fast-moving repricing. Debt-to-GDP has been deteriorating across the developed world for fifteen years, and yields spent most of that period falling. On top of this, foreign holdings of Treasuries hit a record $9.37trn in May 2026, which is not the profile of a buyers' strike. Brad Setser of the Council on Foreign Relations has spent years pointing out that the TIC data everyone quotes ‘only measures China's holdings of Treasuries in US custodians’ so the widely cited collapse in Chinese ownership is largely a change of settlement venue rather than of intent. The number that does bite is the interest bill rather than the debt stock. US net interest reached $857bn between October 2025 and June 2026, up 13% year-on-year, and now exceeds both defence and Medicare. That is a genuine doom loop, but it is a consequence of higher yields as much as a cause of them.

Term premium, and the buyers who left

Term premium, which is the extra yield demanded for locking money up for thirty years rather than rolling short paper, has gone from a rounding error to the dominant driver. The New York Fed's ACM estimate sits around 0.80%, positive for the first time since 2023. That sounds like a complete explanation until you ask what changed, and the honest answer is the buyer base rather than the mood.

Every price-insensitive buyer of long duration has withdrawn at once. The Fed concluded its balance sheet runoff on 1 December 2025 and is buying bills again. UK pension schemes, post-LDI and increasingly bought out, no longer need the ultra-long gilt in the volume they once did. The Bank of Japan is normalising. What has stepped in is leverage: the Dallas Fed puts hedge fund net repo borrowing at roughly $1.8trn, about 6% of marketable notes and bonds, by the end of 2025. The marginal buyer of the world's risk-free asset is now a levered relative-value fund.

However, term premium is a model output rather than something you can observe. The two standard estimates, ACM and Kim-Wright, routinely disagree by a quarter of a percentage point, so the level is a range and not a fact. And QT ending should have been bullish for bonds; yields rose anyway. That is a fair challenge to the plumbing thesis, but it cuts the other way on fragility: a market cleared by levered funds is one where a volatility shock forces selling rather than buying. The BIS documented exactly that transmission during the April 2025 turbulence.

Japan: The world's ATM is closing

Anshul Pradhan of Barclays made the observation that should bother everyone, noting that ‘three separate US economic releases this month pointed toward lower yields yet long-end yields rose anyway’. Part of the explanation sits in Tokyo, and it is mechanical rather than emotional.

For two decades Japanese life insurers exported the country's savings into Treasuries. A Japanese insurer buying a Treasury must hedge the dollar exposure, and the cost of that hedge is roughly the gap between US and Japanese short rates. With US short rates far above Japan's, the hedge now eats almost the entire yield. TD Economics calculates that a Japanese investor can earn around 2.3% at home against closer to 1.3% on a fully hedged Treasury. Ministry of Finance data show net sales of foreign securities of ¥4trn (about $25bn) since the start of 2026. The whale has gone home, and it did so because of arithmetic, not sentiment. However, roughly 90% of JGBs are held domestically, so Japan is self-funded and the JGB move is arguably a technical story within its own insurance industry rather than a fiscal verdict. The carry gap also still favours the dollar: even after the Bank of Japan's December hike to 0.75%, US rates sit around 3.75%. Repatriation is real but choppy, and it is a flow rather than a permanent regime change.

AI capex arrives in the bond market

This is where the last article and this one meet. The hyperscalers have moved from funding capex with operating cash flow to funding it with debt. Five of them issued $159bn of bonds through early June 2026, up 47% year-on-year against $121bn for the whole of 2025, and Morgan Stanley expects roughly $570bn of AI-related issuance globally this year. Al Cattermole of Mirabaud captured the break in the contract, noting that capex ‘was still going to be equity or cash funded’ and that ‘by bringing capex spend into the debt markets, you now have the question of credit worthiness’. Long-dated corporate supply competes directly with government supply for the same finite pool of duration buyers, and bear-steepening is the arithmetic result.

However, the demand has been enormous. Oracle's eight-tranche deal in February drew a peak order book of around $129bn, breaking Meta's $125bn record, and high yield spreads sit near 281bps, in the richest decile on record. If credit were genuinely worried, that is not what it would look like. A word of caution on those order books, because they are quoted far more often than they are understood. In a syndicated deal, investors deliberately pad orders knowing allocations are scaled back pro rata, and books are built at generous initial price talk that tightens as demand comes in. A headline multiple is therefore partly an artefact of the bookbuilding process. The honest tests are the new-issue concession the borrower must pay and how the bonds trade afterwards and Oracle's September 2025 paper subsequently traded some 60-80bps wide of issue despite a record book.

Gold and the safe asset question

At the end of 2025 gold made up 27% of official reserves against 22% for US Treasuries, the first-time gold has outranked Treasuries since 1996. Michael Hartnett of BofA has built a whole framework on it whith the phrase ‘anything but bonds’, and a World Gold Council survey found 74% of central banks expect lower dollar holdings within five years. It reads like the end of an era. However, it is mostly a price effect. The ECB is explicit that the crossover was ‘driven mainly by valuation effects rather than purchases’: the average gold price rose 43% year-on-year in 2025 while official sector buying fell to around 850 tonnes from more than 1,000 a year in 2022-24. Central banks did not swap Treasuries for gold. Gold went up. That is a mark-to-market fact rather than a flow one, and worth remembering before anyone writes the dollar's obituary.

Overview

The five explanations are not equally weighted, and the sequencing matters. The fiscal numbers are the reason the long end is vulnerable, but they are not the reason it broke this month; they have been deteriorating on a steady trend while yields have moved in a step. The proximate cause is a demand vacuum. The Fed, the Japanese life insurer and the UK pension scheme each stepped back from the thirty-year point within roughly eighteen months of one another, and the AI complex chose precisely that window to bring several hundred billion dollars of long-dated corporate paper to market. The replacement bid is levered, which is fine until volatility forces it to sell. Mohamed El-Erian's verdict on the Treasury's buyback that ‘the effects of this financial engineering are short dated unless followed by fundamental policy adjustments’ applies to the episode as a whole. The long end is not pricing a default. It is pricing the absence of anyone who must buy.